Super-resolution ultrasonic imaging method in non-contrast mode

Abstract

The invention provides a super-resolution ultrasonic imaging method in a non-contrast mode, which is characterized by comprising the following steps of: S1, acquiring a plurality of ultrasonic images of blood cells in an imaging target area through a recursive synthetic aperture imaging method; s2, removing motion artifacts of each ultrasonic image according to a speckle tracking method to obtain a corresponding reconstructed ultrasonic image; s3, filtering each reconstructed ultrasonic image according to a space-time filtering method to obtain a corresponding blood cell ultrasonic image; s4, constructing a training data set and a generative adversarial network, and training the generative adversarial network according to the training data set to obtain a blood cell positioning model; s5, inputting each blood cell ultrasonic image into the blood cell positioning model to obtain a corresponding positioning result; and S6, obtaining a super-resolution ultrasonic microscopic image according to all positioning results. In a word, the method can quickly generate the super-resolution ultrasonic microscopic image.

Description

Technical Field

[0001] The present invention relates to an ultrasonic imaging technology, and in particular to a super-resolution ultrasonic imaging method in a non-contrast imaging mode. Background Art

[0002] Compared to other medical imaging technologies, such as X-ray imaging and magnetic resonance imaging, ultrasound imaging has the advantages of being radiation-free, fast, and low-cost, and is now widely used in the biomedical field. However, due to the acoustic diffraction limit, the spatial resolution of ultrasound imaging is limited to half a wavelength.

[0003] To further improve the resolution of ultrasound imaging, ultrasound localization microscopy (ULM) was proposed and applied in pre-clinical trials and clinical disease diagnosis. With the help of ultrasound contrast-enhanced microbubbles and single-molecule localization methods, ULM has successfully broken through the acoustic diffraction limit and achieved ultrasound imaging of tiny tissues at the micron scale. Specifically, in the traditional ULM imaging process, the echoes from the flowing microbubbles in the tissue are collected by an ultrasonic transducer to reconstruct a series of B-mode ultrasound images. The microbubbles are then located frame by frame using a positioning algorithm. The resulting series of positioning frames are superimposed to reconstruct an ultra-high-resolution ULM image.

[0004] However, existing ULM imaging techniques typically require the use of ultrasound contrast microbubbles to enhance ultrasound echo signals. This leads to challenges such as limited application scenarios, slow imaging speed, and limited detection power, significantly limiting the practical application of ULM technology. Specifically, as exogenous imaging markers, ultrasound contrast agents present numerous challenges in experimental settings. Consequently, they are not an internationally accepted auxiliary imaging agent and are approved for use in only a little over 40 countries worldwide. Secondly, when implementing ULM imaging, existing ULM localization methods are typically based on single-molecule localization approaches. The reconstruction quality of such ULM methods depends heavily on the accuracy of microbubble localization. To improve localization accuracy, microbubbles must be as sparse as possible to facilitate the centering of individual microbubbles. Accumulating sufficient microbubble localization information to reconstruct super-resolution images requires sufficient ultrasound data acquisition, which results in lengthy data acquisition times and limits its application in rapid biological activity monitoring, such as hemodynamics and brain activity. In addition, ultrasound contrast microbubbles will rupture under the action of high-power ultrasound, so the mechanical index of the incident ultrasound is limited to below 0.2, which reduces the image signal-to-noise ratio.

[0005] In summary, the existing ULM imaging technology based on ultrasound contrast-enhanced microbubbles still has problems such as limited scope of application and slow imaging speed. Summary of the Invention

[0006] The present invention is made to solve the above problems, and its purpose is to provide a super-resolution ultrasonic imaging method in a non-contrast imaging mode.

[0007] The present invention provides a super-resolution ultrasound imaging method in a non-contrast imaging mode, which has the following characteristics: step S1, using a recursive synthetic aperture imaging method in combination with high-frequency ultrasonic transmission waves to collect multiple ultrasound images of blood cells in the imaging target area; step S2, using a speckle tracking method to remove motion artifacts from each ultrasound image to obtain a corresponding reconstructed ultrasound image; step S3, using a spatiotemporal filtering method to filter each reconstructed ultrasound image to obtain a corresponding blood cell ultrasound image; step S4, constructing a training data set and a generative adversarial network, and training the generative adversarial network based on the training data set to obtain a blood cell positioning model; step S5, inputting each blood cell ultrasound image into the blood cell positioning model to obtain a corresponding positioning model. Positioning result; Step S6, obtaining a super-resolution ultrasound microscopy image based on all positioning results, wherein Step S4 includes the following sub-steps: Step S4-1, randomly placing different numbers of blood cells in a grid area of ​​a fixed size to generate multiple blood cell images; Step S4-2, performing ultrasound image simulation on each blood cell image in turn according to the PSF convolution method to obtain a corresponding simulated blood cell ultrasound image; Step S4-3, using all the simulated blood cell ultrasound images as a training data set; Step S4-4, using the initial blood cell positioning model as a generator and PatchGAN as a discriminator to construct a generative adversarial network; Step S4-5, training the generative adversarial network according to the training data set to obtain a trained generator as the blood cell positioning model.

[0008] The super-resolution ultrasound imaging method in the non-contrast mode provided by the present invention may further have the following features: wherein, in step S4-2, the expression for obtaining the simulated blood cell ultrasound image according to the PSF convolution method is: B(x,y)=||r(x,y)+jH[r(x,y)]||, r(x,y)=ξ(x,y)*h(x,y), Where r(x,y) is the real position of a single blood cell in the blood cell image, H is the Hilbert transform, B(x,y) is the simulated blood cell ultrasound image corresponding to the blood cell, h(x,y) is the system PSF corresponding to the blood cell, x is the ultrasound propagation direction, y is the imaging plane direction, c is the sound speed, f is the emission center frequency, σ x and σ y are the system PSF parameters in the ultrasound propagation direction and imaging plane direction, respectively.

[0009] The super-resolution ultrasound imaging method in the non-contrast mode provided by the present invention may also have the following characteristics: wherein the σ corresponding to each blood cell is randomly set x and σy ,σ x ∈[77~155μm],σ y ∈[55~110μm], c=1540m / s, f=15.625MHz, and the pixel size is 55μm.

[0010] The super-resolution ultrasound imaging method in the non-contrast mode provided by the present invention may also have the following features: wherein, in step S4-5, the Adam optimization gradient descent algorithm is combined with the learning rate decay strategy to perform multiple rounds of training on the generative adversarial network until the learning rate is 0, thereby obtaining a blood cell localization model, and the discriminator and the generator are iteratively updated according to the corresponding losses in each round of training.

[0011] The super-resolution ultrasound imaging method in the non-contrast mode provided by the present invention may also have the following characteristics: wherein the loss of the generator is expressed as: L G =α×L1(G(x),y)-β×logD(G(x),x), where L G is the loss of the generator, L1() is the L1 regularization loss, x is the simulated blood cell ultrasound image input to the generator, G(x) is the image generated by the generator based on the simulated blood cell ultrasound image x, y is the true position of the blood cells in the simulated blood cell ultrasound image x, logD() is the generation adversarial loss, D() is the discriminator output value, α and β are the weights of the L1 regularization loss and the generation adversarial loss respectively, and the expression of the L1 regularization loss is: Where N is the number of pixels.

[0012] The super-resolution ultrasound imaging method in the non-contrast mode provided by the present invention may also have the following characteristics: wherein the loss expression of the discriminator is: L D = -logD(y, x) - log(1-D(G(x), x)), where L D is the loss of the discriminator, x is the simulated blood cell ultrasound image input to the generator, G(x) is the image generated by the generator based on the simulated blood cell ultrasound image x, y is the true position of the blood cells in the simulated blood cell ultrasound image x, logD() is the generated adversarial loss, and D() is the discriminator output value.

[0013] In the super-resolution ultrasound imaging method in the non-contrast mode provided by the present invention, it can also have the following characteristics: wherein, the generator is a ResNet network composed of 6 residual modules, each residual module consists of 2 convolutional layers with 3×3 convolution kernels, 2 batch normalization layers, a ReLU nonlinear activation function and a jump connection.

[0014] Functions and effects of the invention

[0015] The super-resolution ultrasound imaging method in a non-contrast imaging mode of the present invention has the following advantages: first, the imaging of super-resolution ultrasound microscopy images using ultrasound images of blood cells has the characteristics of a wide range of applications and fast imaging speed compared to ultrasound contrast imaging microbubbles; second, ultrasound imaging of blood cells using a recursive synthetic aperture imaging method, a speckle tracking method, and a spatiotemporal filtering method can obtain clearer ultrasound images of blood cells; and third, a simulated ultrasound image of blood cells is constructed using a PSF convolution method to train a generative adversarial network, thereby obtaining a blood cell localization model with a high accuracy for obtaining blood cell localization information. Therefore, the super-resolution ultrasound imaging method in a non-contrast imaging mode of the present invention can quickly generate super-resolution ultrasound microscopy images. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 1 is a flow chart of a super-resolution ultrasound imaging method in a non-contrast imaging mode according to an embodiment of the present invention;

[0017] Figure 2 is a schematic structural diagram of a residual module in an embodiment of the present invention;

[0018] Figure 3 Schematic diagram of the principle of training a generative adversarial network in an embodiment of the present invention;

[0019] Figure 4 Schematic diagram of blood cell ultrasound images and positioning results in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the following embodiments and accompanying drawings specifically illustrate the super-resolution ultrasound imaging method in the non-contrast imaging mode of the present invention.

[0021] Figure 1 4 is a flow chart of a super-resolution ultrasound imaging method in a non-contrast imaging mode according to an embodiment of the present invention.

[0022] like Figure 1 As shown, the super-resolution ultrasound imaging method in the non-contrast imaging mode of this embodiment includes the following steps:

[0023] In step S1, multiple ultrasound images of blood cells in the target imaging area are acquired using a recursive synthetic aperture imaging method in combination with high-frequency ultrasound transmission waves. In this embodiment, the recursive synthetic aperture imaging method is used to delay and superimpose radio frequency data to achieve beamforming processing, thereby obtaining a B-mode ultrasound image, i.e., an ultrasound image.

[0024] In step S2, motion artifacts are removed from each ultrasound image using a speckle tracking method to obtain a corresponding reconstructed ultrasound image. In this embodiment, the speckle tracking method is used to estimate the overall motion trend of the ultrasound image and align all frames, i.e., ultrasound images, to a reference frame based on the motion field, thereby suppressing motion artifacts.

[0025] In step S3, each reconstructed ultrasound image is filtered using a spatiotemporal filtering method to obtain a corresponding blood cell ultrasound image. In this embodiment, a spatiotemporal filtering method based on singular value decomposition is used to suppress echo signals, separate blood cells from the imaging background, and obtain a series of blood cell ultrasound images after background noise filtering.

[0026] Step S4: construct a training data set and a generative adversarial network, and train the generative adversarial network based on the training data set to obtain a blood cell positioning model.

[0027] Wherein, step S4 includes the following sub-steps:

[0028] In step S4-1, different numbers of blood cells are randomly placed in a grid area of ​​a fixed size to generate multiple blood cell images. In this embodiment, the fixed size of the grid area is 512*512 pixels.

[0029] In step S4-2, ultrasound image simulation is performed on each blood cell image in sequence using the PSF convolution method to obtain a corresponding simulated blood cell ultrasound image. In this embodiment, Gaussian noise with a signal-to-noise ratio of 30 dB is added to the simulated blood cell ultrasound image to improve the robustness of the generative adversarial network training.

[0030] In step S4-2, the expression of the simulated blood cell ultrasound image obtained according to the PSF convolution method is:

[0031] B(x,y)=||r(x,y)+jH[r(x,y)]||,

[0032] r(x,y)=ξ(x,y)*h(x,y),

[0033]

[0034] Where r(x, y) is the real position of a single blood cell in the blood cell image, H is the Hilbert transform, B(x, y) is the simulated blood cell ultrasound image corresponding to the blood cell, h(x, y) is the system PSF corresponding to the blood cell, x is the ultrasound propagation direction, y is the imaging plane direction, c is the sound speed, f is the emission center frequency, σ x and σ yare the system PSF parameters in the ultrasound propagation direction and the imaging plane direction, respectively. In this embodiment, c = 1540 m / s, f = 15.625 MHz, and the pixel size is 55 μm. In order to make the simulated blood cell ultrasound image more consistent with the subtle size differences between multiple real flowing blood cells, the σ corresponding to each blood cell is randomly set. x and σ y , σ x ∈[77~155μm],σ y ∈[55~110μm].

[0035] Step S4-3: All simulated blood cell ultrasound images are used as a training data set. In this embodiment, the training data set includes 4,000 simulated blood cell ultrasound images and their corresponding real positions of blood cells.

[0036] In step S4-4, a generative adversarial network is constructed using the initial blood cell localization model as the generator and PatchGAN as the discriminator. In this embodiment, PatchGAN divides the input image into multiple sub-regions to perform local scoring of image authenticity, averaging the scores as the global score. This serves as an adversarial loss for the generator update, thereby improving the ability to discern local subtle features and effectively improving the accuracy of positioning recognition.

[0037] Among them, the generator is a ResNet network consisting of 6 residual modules connected sequentially.

[0038] Figure 2 2 is a schematic structural diagram of a residual module in an embodiment of the present invention.

[0039] like Figure 2 As shown, the residual module 10 includes a first convolutional layer 101, a first batch of normalization layers 102, a ReLU nonlinear activation function 103, a second convolutional layer 104, a second batch of normalization layers 105 and a feature fusion layer 106, which are sequentially connected, i.e., a jump connection.

[0040] The feature fusion layer 106 is used to fuse the features input to the residual module 10 with the features output by the second batch normalization layer 105 to obtain the output features of the residual module 10.

[0041] Step S4-5: Train the generative adversarial network according to the training data set to obtain a trained generator as the blood cell positioning model.

[0042] Figure 3 2 is a schematic diagram of the principle of training a generative adversarial network in an embodiment of the present invention.

[0043] like Figure 3As shown, the simulated blood cell ultrasound image in the training data set is input into the generator to obtain the corresponding positioning recognition result. The discriminator obtains the discriminator output value based on the input positioning recognition result and the actual position of the blood cells corresponding to the simulated blood cell ultrasound image, thereby constructing the generator loss and discriminator loss. In each round of training, the generator is updated by the generator loss, and the discriminator is updated by the discriminator loss.

[0044] In this embodiment, the Adam gradient descent algorithm is optimized, and 100 cycles of training are performed, that is, the training data set is traversed 100 times, the batch size is set to 12, the patch size of the discriminator is set to 70×70, and the initial learning rate is 1×10 -4 To minimize the loss function as quickly as possible, a learning rate decay strategy was used to gradually reduce the learning rate to 0, gradually approaching the optimal parameters. Ultimately, it took 2.5 hours to complete the training of the generative adversarial network and obtain a blood cell localization model.

[0045] The loss of the generator is expressed as follows:

[0046] L G =α×L1(G(x),y)-β×logD(G(x),x),

[0047] Where L G is the loss of the generator, L1() is the L1 regularization loss, x is the simulated blood cell ultrasound image input to the generator, G(x) is the image generated by the generator based on the simulated blood cell ultrasound image x, y is the true position of the blood cells in the simulated blood cell ultrasound image x, logD() is the generated adversarial loss, D() is the discriminator output value, α and β are the weights of the L1 regularization loss and the generated adversarial loss, respectively.

[0048] The expression of L1 regularization loss is:

[0049]

[0050] Where N is the number of pixels.

[0051] The loss of the discriminator is expressed as:

[0052] L D =-logD(y,x)-log(1-D(G(x),x)),

[0053] Where L Dis the loss of the discriminator, x is the simulated blood cell ultrasound image input to the generator, G(x) is the image generated by the generator based on the simulated blood cell ultrasound image x, that is, the positioning recognition result, y is the true position of the blood cells in the simulated blood cell ultrasound image x, logD() is the generated adversarial loss, and D() is the discriminator output value.

[0054] Step S5: Input each blood cell ultrasound image into the blood cell positioning model to obtain a corresponding positioning result.

[0055] Figure 4 Schematic diagram of blood cell ultrasound images and positioning results in an embodiment of the present invention.

[0056] like Figure 4 As shown in the figure, (a) is a blood cell ultrasound image, (b) is the positioning result corresponding to the blood cell ultrasound image, (c) is an enlarged image of the white rectangular area in the blood cell ultrasound image, and (d) is an enlarged image of the white rectangular area in the positioning result. The white dots in (c) are blood cells, the centers of the white circles in (d) are the actual positions of the corresponding blood cells in (c), and the white dots are the position recognition results of the blood cell positioning model for the corresponding blood cells. This shows that the blood cell positioning model can accurately identify the position of blood cells and obtain accurate blood cell position information.

[0057] Step S6: Obtain a super-resolution ultrasound microscopy image based on all positioning results.

[0058] The super-resolution ultrasound imaging method in a non-contrast mode provided in this embodiment can obtain super-resolution ultrasound microscopic images by accurately locating high-concentration blood cells without the aid of ultrasound contrast agents, thereby improving the flexibility of ultrasound positioning microscopy technology and expanding the scope of application of this method. At the same time, because blood cells are densely distributed in blood vessels, the number of frames required to accumulate sufficient positioning information is greatly reduced. Therefore, compared with existing microbubble imaging technology, this method can achieve image reconstruction based on a small amount of raw ultrasound data, speed up imaging, and thus improve the temporal resolution of imaging. In addition, based on this method, parameter dependence is avoided in the positioning process, and no additional manual intervention is required.

[0059] Functions and Effects of the Embodiments

[0060] According to the non-contrast imaging mode super-resolution ultrasound imaging method involved in this embodiment, first, the ultrasound image of blood cells is used to generate super-resolution ultrasound microscopic images, which has the characteristics of a wider range of applications and faster imaging speed compared to ultrasound contrast microbubbles. Second, the ultrasound imaging of blood cells is performed using recursive synthetic aperture imaging, speckle tracking, and spatiotemporal filtering methods, which can obtain clearer blood cell ultrasound images. Third, the PSF convolution method is used to construct simulated blood cell ultrasound images for training the generative adversarial network, thereby obtaining a blood cell localization model with high accuracy for obtaining blood cell localization information. In summary, this method can quickly generate super-resolution ultrasound microscopic images.

[0061] Those skilled in the art will appreciate that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A super-resolution ultrasound imaging method in a non-contrast imaging mode, characterized in that: The following steps are involved: Step S1, using a recursive synthetic aperture imaging method in combination with high-frequency ultrasonic transmission waves, to acquire multiple ultrasonic images of blood cells in the imaging target area; Step S2, removing motion artifacts from each of the ultrasound images according to a speckle tracking method to obtain a corresponding reconstructed ultrasound image; Step S3, filtering each of the reconstructed ultrasound images according to a spatiotemporal filtering method to obtain a corresponding blood cell ultrasound image; Step S4, constructing a training data set and a generative adversarial network, and training the generative adversarial network according to the training data set to obtain a blood cell positioning model; Step S5, inputting each of the blood cell ultrasound images into the blood cell positioning model to obtain corresponding positioning results; Step S6: obtaining a super-resolution ultrasound microscopic image based on all the positioning results. Wherein, the step S4 includes the following sub-steps: Step S4-1, randomly placing different numbers of blood cells in a grid area of ​​a fixed size to generate multiple blood cell images; Step S4-2, performing ultrasound image simulation on each of the blood cell images in sequence according to the PSF convolution method to obtain a corresponding simulated blood cell ultrasound image; Step S4-3, using all the simulated blood cell ultrasound images as the training data set; Step S4-4, using the initial blood cell positioning model as a generator and PatchGAN as a discriminator to construct the generative adversarial network; Step S4-5: training the generative adversarial network according to the training data set to obtain the trained generator as the blood cell positioning model.

2. The super-resolution ultrasound imaging method in non-contrast imaging mode according to claim 1, characterized in that: in, In step S4-2, the expression for the simulated blood cell ultrasonic image obtained according to the PSF convolution method is: B(x, y)=||r(x, y)+jH[r(x, y)]||, r(x,y)=ξ(x,y)*h(x,y), Where r(x, y) is the real position of a single blood cell in the blood cell image, H is the Hilbert transform, B(x, y) is the simulated blood cell ultrasound image corresponding to the blood cell, h(x, y) is the system PSF corresponding to the blood cell, x is the ultrasound propagation direction, y is the imaging plane direction, c is the sound speed, f is the emission center frequency, σ x and σ y are the system PSF parameters in the ultrasound propagation direction and imaging plane direction, respectively.

3. The super-resolution ultrasound imaging method in non-contrast imaging mode according to claim 2, characterized in that: in, Randomly set the σ corresponding to each blood cell x and σ y , σ x ∈[77~155μm],σ y ∈[55~110μm], c = 1540 m / s, f = 15.625 MHz, and pixel size is 55 μm.

4. The super-resolution ultrasound imaging method in non-contrast mode according to claim 1, characterized in that: in, In step S4-5, the generative adversarial network is trained for multiple rounds by using the Adam optimized gradient descent algorithm and combining it with a learning rate decay strategy until the learning rate reaches 0, thereby obtaining the blood cell positioning model. The discriminator and the generator are iteratively updated according to the corresponding losses in each round of training.

5. The super-resolution ultrasound imaging method in non-contrast imaging mode according to claim 4, characterized in that: in, The loss expression of the generator is: L G =α×L1(G(x),y)-β×logD(G(x),x), Where L G is the loss of the generator, L1() is the L1 regularization loss, x is the simulated blood cell ultrasound image input to the generator, G(x) is the image generated by the generator based on the simulated blood cell ultrasound image x, y is the true position of the blood cells in the simulated blood cell ultrasound image x, logD() is the generation adversarial loss, D() is the discriminator output value, α and β are the weights of the L1 regularization loss and the generation adversarial loss, respectively. The expression of the L1 regularization loss is: Where N is the number of pixels.

6. The super-resolution ultrasound imaging method in non-contrast mode according to claim 4, characterized in that: in, The loss expression of the discriminator is: L D = -log D(y, x) - log(1 - D(G(x), x)), Where L D is the loss of the discriminator, x is the simulated blood cell ultrasound image input to the generator, G(x) is the image generated by the generator based on the simulated blood cell ultrasound image x, y is the true position of the blood cells in the simulated blood cell ultrasound image x, logD() is the generated adversarial loss, and D() is the discriminator output value.

7. The super-resolution ultrasound imaging method in non-contrast mode according to claim 1, characterized in that: in, The generator is a ResNet network composed of 6 residual modules. Each residual module consists of two convolutional layers with 3×3 convolution kernels, two batch normalization layers, a ReLU nonlinear activation function, and a skip connection.

Images